DEV Community

Aman
Aman

Posted on Edited on

Building an Active Circuit Breaker for the Agent Economy on Google ADK & AP2

#ai #programming #googlecloud

Disclaimer: I created this project and article for the purposes of entering the Google All Things Agentic Hackathon on Devpost.

1. Introduction: The Unsupervised Agent Economy

In 2026, AI agents are no longer passive chat assistants. Using Google and Coinbase's new Agent Payment Protocol (AP2) and on-chain micropayment standard, agents autonomously procure compute, pay for vector embeddings, and hire specialized sub-agents.

However, giving autonomous background agents financial wallets introduces a critical systemic risk: Runaway Spend Loops and Blast-Radius Acceleration. If an agent encounters an unhandled prompt exception or recursive loop, it can burn through hundreds of dollars in automated micropayments in minutes before any human notices.

Existing API gateways only show post-facto logs after money has already been spent.

To solve this, we built Wardstone AP2 — an active, predictive circuit breaker for the AI Agent Fleet Controller that calculates blast-radius variance and quarantines anomalous payment mandates before on-chain finality on Ethereum Sepolia.

2. Architecture: The Multi-Agent Nexus Pattern

Wardstone is built with Google Agent Development Kit (ADK 2.7) and Gemini 3.5 Flash using a 4-agent separation of concerns:

┌────────────────────────────────────────────────────────────────┐
│              ROOT ORCHESTRATOR (Google ADK)                    │
│      Supervises state machine, retries, and recovery           │
└───────┬────────────────┬────────────────┬───────────────┬──────┘
        │                │                │               │
   ┌────┴─────┐     ┌────┴─────┐     ┌────┴─────┐    ┌────┴─────┐
   │ Watcher  │     │Forecaster│     │Gatekeeper│    │ Forensics│
   │  Agent   │───> │  Agent   │───> │  Agent   │──> │  Agent   │
   └────┬─────┘     └────┬─────┘     └────┬─────┘    └────┬─────┘
        │                │                │               │
   Cloud Pub/Sub  Memory Bank    Ethereum Sepolia Gemini 3.5
     Ingestion    (Baselines)     Settlement     Postmortems
Enter fullscreen mode Exit fullscreen mode

1. Watcher Agent

Subscribes to Google Cloud Pub/Sub (mandate-events), normalizes AP2 payloads (Intent, Cart, Payment Mandate), and writes records to Google Cloud Firestore.

2. Forecaster Agent

Queries the Agent Engine Memory Bank for the agent's historical spend profile. Calculates moving-window rolling velocity and assigns a dynamic 0-100 Blast-Radius Risk Score. Crucially, it monitors the destination burst tracker to catch cross-agent collusion—detecting when multiple distinct agents attempt to sybil-attack or smurf funds into a single target wallet.

3. Gatekeeper Agent

Enforces governance thresholds:

  • Score < 60 (Low Risk): Authorizes official on-chain settlement on Ethereum Sepolia testnet, generating a verifiable transaction hash on Etherscan.
  • Score >= 60 (High Risk): Trips the circuit breaker, halting settlement with provably zero on-chain token movement.
  • Exposes an official A2A Agent Card (JSON-LD) allowing external buyer agents to query pre-clearance prior to signing.

4. Forensics Agent (Powered by Gemini 3.5 Flash)

When a mandate is quarantined, the Forensics Agent prompts Gemini 3.5 Flash to generate a structured, executive-ready incident postmortem detailing the statistical deviation, affected components, and recommended remediation. It also implements an Institutional Memory Feedback Loop—querying past human overrides (FORCE_APPROVE) so the LLM natively learns from past governance actions and injects that context into future anomaly explanations.

3. Failure-Tolerant Multi-Agent Recovery

Under the hackathon's Multi-Agent Nexus architectural criteria, our system implements resilient failure recovery:
If a worker agent crashes or returns malformed schema, the FailureRecoveryManager detects the exception, attempts exponential retries, and applies safe defensive fallback quarantine without crashing the fleet.

4. Key Learnings & Takeaways

  1. Active Gating > Passive Observability: You cannot rely on post-execution dashboards when autonomous agents have real wallets. Control planes must act as pre-settlement immune systems.
  2. A2A Agent Cards Enable Collaborative Governance: Exposing pre-clearance query RPCs allows well-behaved agents to verify their spend headroom before committing expensive transactions.

Built with Google ADK, Gemini 3.5 Flash, Google Cloud Run, Cloud Firestore, Cloud Pub/Sub, and Ethereum Sepolia.

Top comments (0)